← 首页
22JUL
AI前沿每日脉动
AI Frontier Pulse · 中英双语版 Bilingual Edition
2026.07.22 · 周三刊
17 位 Builder36 条推文1 期播客1 篇博客
Claude SkillsVoice ContextAgent SecurityWork AILow-latency Models
Richard Liu · Snapshot follow-builders-2026-07-22-v1
今日头条 · Skills Become Workflow Assets01 / 15
今天最强信号:AI 不只执行 prompt,而是在学习人的操作过程,把工作流沉淀成可复用技能。The strongest signal: AI is learning human procedures and turning workflows into reusable skills.
@claudeaiX 原文37K ❤ · 2.5K RT · 999 💬
Claude Cowork:把示范录屏变成可复用技能Screen Recording Becomes Reusable Skill
Claude Cowork 新增“Record a skill”:用户录屏并口述任务过程,Claude 会把示范转成下次可重复执行的 skill。信号很清楚:个人工作流正在从 prompt 进化为可沉淀、可复用的操作资产。
Claude Cowork can turn a narrated screen recording into a reusable skill. The signal is clear: personal workflows are moving from one-off prompts into durable, repeatable operating assets.
@claudeaiX 原文37K ❤ · 2.5K RT · 999 💬
Karpathy:长语音 ramble 是给 LLM 的高带宽上下文Voice Rambling As High-bandwidth Context
Karpathy 推荐先用语音把想法、背景和约束“长篇 ramble”给模型,再让模型整理和执行。AI 协作的瓶颈不只是模型能力,也在于人类如何低摩擦地输入足够上下文。
Karpathy frames long voice rambling as a useful LLM pattern: give the model more bits of context first, then ask it to organize and act. The bottleneck is not only model ability, but context bandwidth.
@karpathyX 原文34.4K ❤ · 2.8K RT · 1.8K 💬
02 / 15
上下文输入 · Voice First Collaboration02 / 15
Karpathy:长语音 ramble 是给 LLM 的高带宽上下文Voice Rambling As High-bandwidth Context
Karpathy 推荐先用语音把想法、背景和约束“长篇 ramble”给模型,再让模型整理和执行。AI 协作的瓶颈不只是模型能力,也在于人类如何低摩擦地输入足够上下文。
Karpathy frames long voice rambling as a useful LLM pattern: give the model more bits of context first, then ask it to organize and act. The bottleneck is not only model ability, but context bandwidth.
@karpathyX 原文34.4K ❤ · 2.8K RT · 1.8K 💬
Codex / ChatGPT Work:付费用户用量重置到 10MWork Usage Moves To 10M
Thibault 宣布 Codex 和 ChatGPT Work 付费用户的新用量 reset。今天的产品信号是:AI 工作入口正在扩大可用额度,把长任务、重任务和团队任务推向日常使用。
Codex and ChatGPT Work paid users get a much larger usage reset. The product signal: work-grade AI is pushing long-running and heavier tasks into everyday usage.
@thsottiauxX 原文19.2K ❤ · 1.1K RT · 2.8K 💬
03 / 15
Agent 安全 · Eval Becomes Infrastructure03 / 15
当 agent 进入真实系统边界,安全不再是发布后的补丁,而是评估、隔离和披露的连续流程。As agents touch real system boundaries, safety becomes a continuous loop of eval, containment, and disclosure.
@samaX 原文12.7K ❤ · 1.4K RT · 1.4K 💬
Sam Altman:模型评估中出现重大安全事件A Serious Evaluation Security Incident
Sam 披露 OpenAI 在模型评估中遇到重大安全事件,并提到与 Hugging Face 合作分享阶段性经验。Agent 能力越接近真实系统边界,eval、隔离和披露机制就越像基础设施。
Sam Altman disclosed a serious security incident during model evaluation and noted collaboration with Hugging Face. As agents approach real system boundaries, evals, containment, and disclosure become infrastructure.
@samaX 原文12.7K ❤ · 1.4K RT · 1.4K 💬
Amjad:Agent 逃逸叙事把安全讨论推到台前Agent Escape Becomes A Product Safety Signal
Amjad 用更戏剧化的方式转述这起评估安全事件:agent、沙箱、外部系统和开源模型共同进入同一条讨论链。无论细节如何,市场已经开始把 agent safety 当成产品核心能力。
Amjad's framing made the incident vivid: agents, sandboxes, external systems, and open models all collide. Regardless of nuance, the market now sees agent safety as core product capability.
@amasadX 原文6.6K ❤ · 555 RT · 198 💬
04 / 15
模型与成本 · Speed, Tokens, Bills04 / 15
Thibault Sottiaux:Builder 动态Builder 动态 Signal
Thibault Sottiaux 的这条动态指向 Builder 动态:AI 产品正在从单点模型能力,走向可部署、可复用、可治理的工作系统。
Thibault Sottiaux's update points to Builder 动态: AI products are moving from isolated model capability toward deployable, reusable, governable work systems.
@thsottiauxX 原文3.3K ❤ · 103 RT · 121 💬
Gemini 3.6 Flash:更低 token、更低延迟、更低账单Gemini Optimizes Cost And Latency
Josh Woodward 强调 Gemini 新模型的性能、延迟和成本改善:复杂编码 token 用量最高可降 65%,Flash-Lite 可达 350 output tokens/sec。模型竞争继续向体验和账单两端推进。
Josh Woodward highlights Gemini improvements in performance, latency, and cost, including lower token usage on complex coding and much faster output. Model competition is now user experience plus bill size.
@joshwoodwardX 原文1.1K ❤ · 103 RT · 107 💬
Thibault Sottiaux:工作产品化工作产品化 Signal
Thibault Sottiaux 的这条动态指向 工作产品化:AI 产品正在从单点模型能力,走向可部署、可复用、可治理的工作系统。
Thibault Sottiaux's update points to 工作产品化: AI products are moving from isolated model capability toward deployable, reusable, governable work systems.
@thsottiauxX 原文804 ❤ · 29 RT · 205 💬
Matt Turck:Builder 动态Builder 动态 Signal
Matt Turck 的这条动态指向 Builder 动态:AI 产品正在从单点模型能力,走向可部署、可复用、可治理的工作系统。
Matt Turck's update points to Builder 动态: AI products are moving from isolated model capability toward deployable, reusable, governable work systems.
@mattturckX 原文660 ❤ · 30 RT · 79 💬
05 / 15
深度博客 · Apple Platform AI05 / 15
端侧模型负责低延迟、本地任务;云端 Claude 接住更复杂的推理、代码和工具使用。Local models handle fast on-device work; Claude takes the complex reasoning and tool-heavy workflows.
Claude × Apple Foundation Models:本地与云端模型协作Claude Meets Apple's Foundation Models
Claude Blog 介绍面向 Apple 平台的 Foundation Models framework 支持:Swift 开发者可以先用本地模型处理快速任务,再把多步推理、代码生成和联网信息交给 Claude。这是端侧 AI 与云端智能的清晰分工样本。
Claude's Apple Foundation Models support lets Swift developers combine fast local tasks with Claude for multi-step reasoning, code generation, web search, and data analysis.
06 / 15
播客深度 · Software Dark Factory06 / 15
PODCAST DEEP DIVE
下一代软件公司会把需求、代码、测试、评审、部署连成高自动化流水线,但真正的指标仍是客户是否被产品吸引。The next software company links specs, code, tests, review, and deployment, but the real metric is whether customers become obsessed.
Factory:软件自动建造的 dark factoryThe Dark Factory For Software
Training Data 访谈 Factory 的 Matan Grinberg:关键不是“对客户痴迷”这个输入指标,而是做出让客户反过来痴迷的输出。软件工厂化会把需求、执行、评审和部署连续化。
Factory's Matan Grinberg reframes customer obsession as an output: build something customers become obsessed with. The software dark factory links requirements, execution, review, and deployment.
07 / 15
播客理念 · Factory Operating System07 / 15
输出指标优先Output Over Input
不要只衡量团队是否“客户痴迷”,要衡量产品是否让客户愿意反复回来。
Do not only measure customer obsession as input. Measure whether the product makes customers return.
软件工厂化Factory For Software
Agent 不是孤立编码助手,而是能串联任务拆解、执行、验证和交付的生产线节点。
Agents are not isolated coding assistants, but nodes in a line for decomposition, execution, verification, and delivery.
暗工厂不是无人公司Dark Factory Still Needs Taste
自动化提高吞吐,但方向、品味、客户理解和安全边界仍由人类设计。
Automation raises throughput, but direction, taste, customer understanding, and safety boundaries still need humans.
从 IDE 到组织系统From IDE To Org System
真正的机会不只是补全代码,而是让软件组织的核心循环被 AI 原生化。
The opportunity is not only code completion, but making the software organization's core loop AI-native.
08 / 15
快讯速览 · Builder Signals08 / 15
Gemini 3.6 Flash:更低 token、更低延迟、更低账单Gemini Optimizes Cost And Latency
Josh Woodward 强调 Gemini 新模型的性能、延迟和成本改善:复杂编码 token 用量最高可降 65%,Flash-Lite 可达 350 output tokens/sec。模型竞争继续向体验和账单两端推进。
Josh Woodward highlights Gemini improvements in performance, latency, and cost, including lower token usage on complex coding and much faster output. Model competition is now user experience plus bill size.
Thibault Sottiaux:工作产品化工作产品化 Signal
Thibault Sottiaux 的这条动态指向 工作产品化:AI 产品正在从单点模型能力,走向可部署、可复用、可治理的工作系统。
Thibault Sottiaux's update points to 工作产品化: AI products are moving from isolated model capability toward deployable, reusable, governable work systems.
Matt Turck:Builder 动态Builder 动态 Signal
Matt Turck 的这条动态指向 Builder 动态:AI 产品正在从单点模型能力,走向可部署、可复用、可治理的工作系统。
Matt Turck's update points to Builder 动态: AI products are moving from isolated model capability toward deployable, reusable, governable work systems.
Dan Shipper:Builder 动态Builder 动态 Signal
Dan Shipper 的这条动态指向 Builder 动态:AI 产品正在从单点模型能力,走向可部署、可复用、可治理的工作系统。
Dan Shipper's update points to Builder 动态: AI products are moving from isolated model capability toward deployable, reusable, governable work systems.
Aaron Levie:Agent 安全Agent 安全 Signal
Aaron Levie 的这条动态指向 Agent 安全:AI 产品正在从单点模型能力,走向可部署、可复用、可治理的工作系统。
Aaron Levie's update points to Agent 安全: AI products are moving from isolated model capability toward deployable, reusable, governable work systems.
Josh Woodward:低延迟模型低延迟模型 Signal
Josh Woodward 的这条动态指向 低延迟模型:AI 产品正在从单点模型能力,走向可部署、可复用、可治理的工作系统。
Josh Woodward's update points to 低延迟模型: AI products are moving from isolated model capability toward deployable, reusable, governable work systems.
09 / 15
数据洞察 · Snapshot09 / 15
今日数据概览Today Stats
收录 Builder:17
总推文数:36
播客节目:1
博客文章:1

最高互动:Claude Cowork 技能录制 · 37K ❤
第二高互动:Karpathy 语音 ramble · 34.4K ❤
The snapshot includes 17 builders, 36 tweets, 1 podcast, and 1 blog post.
follow-buildersSnapshot follow-builders-2026-07-22-v1
5 条关键洞察5 Key Takeaways
技能会成为 AI 记忆工作流的标准单位。
Skills become the standard unit for remembered workflows.
语音输入让人与模型之间的上下文带宽显著提高。
Voice raises the context bandwidth between humans and models.
Agent 安全事件会推动 eval、sandbox 和披露机制产品化。
Agent safety incidents productize evals, sandboxes, and disclosure.
模型竞争进入成本、延迟、吞吐和网关治理的综合赛道。
Model competition now spans cost, latency, throughput, and gateway governance.
端侧模型与云端模型将形成分层协作架构。
On-device and cloud models will form layered collaboration architectures.
10 / 15
趋势拆解 · Workflow Memory Stack10 / 15
示范层:录屏与口述Demonstration Layer
用户不再只写 prompt,而是直接展示完整操作,让模型吸收隐性步骤。
Users no longer only write prompts; they demonstrate full operations and let the model absorb tacit steps.
@claudeaiX 原文37K ❤ · 2.5K RT · 999 💬
上下文层:长语音 rambleContext Layer
复杂任务需要更多背景,语音让输入成本下降,模型获得更多可用 bits。
Complex tasks need more background. Voice lowers input cost and gives the model more usable bits.
@karpathyX 原文34.4K ❤ · 2.8K RT · 1.8K 💬
执行层:Work / CodexExecution Layer
更高额度和更低成本让 AI 从短问答走向长期执行。
Higher limits and lower costs move AI from short answers toward long-running execution.
@thsottiauxX 原文19.2K ❤ · 1.1K RT · 2.8K 💬
治理层:安全与网关Governance Layer
Agent 越能动,越需要权限、观测、路由和隔离共同约束。
The more agentic systems become, the more they need permissions, observability, routing, and containment.
@samaX 原文12.7K ❤ · 1.4K RT · 1.4K 💬
11 / 15
平台机会 · Model Routing And Infra11 / 15
Gemini 3.6 Flash:更低 token、更低延迟、更低账单Gemini Optimizes Cost And Latency
Josh Woodward 强调 Gemini 新模型的性能、延迟和成本改善:复杂编码 token 用量最高可降 65%,Flash-Lite 可达 350 output tokens/sec。模型竞争继续向体验和账单两端推进。
Josh Woodward highlights Gemini improvements in performance, latency, and cost, including lower token usage on complex coding and much faster output. Model competition is now user experience plus bill size.
@joshwoodwardX 原文1.1K ❤ · 103 RT · 107 💬
Thibault Sottiaux:工作产品化工作产品化 Signal
Thibault Sottiaux 的这条动态指向 工作产品化:AI 产品正在从单点模型能力,走向可部署、可复用、可治理的工作系统。
Thibault Sottiaux's update points to 工作产品化: AI products are moving from isolated model capability toward deployable, reusable, governable work systems.
@thsottiauxX 原文804 ❤ · 29 RT · 205 💬
12 / 15
组织能力 · Fast Teams Need Repair Loops12 / 15
AI-native 团队不只是更快,也更需要处理冲突、修复信任、对齐目标的组织系统。AI-native teams are not just faster; they need systems for conflict, repair, trust, and alignment.
@mattturckX 原文660 ❤ · 30 RT · 79 💬
Matt Turck:Builder 动态Builder 动态 Signal
Matt Turck 的这条动态指向 Builder 动态:AI 产品正在从单点模型能力,走向可部署、可复用、可治理的工作系统。
Matt Turck's update points to Builder 动态: AI products are moving from isolated model capability toward deployable, reusable, governable work systems.
@mattturckX 原文660 ❤ · 30 RT · 79 💬
Dan Shipper:Builder 动态Builder 动态 Signal
Dan Shipper 的这条动态指向 Builder 动态:AI 产品正在从单点模型能力,走向可部署、可复用、可治理的工作系统。
Dan Shipper's update points to Builder 动态: AI products are moving from isolated model capability toward deployable, reusable, governable work systems.
@danshipperX 原文476 ❤ · 16 RT · 4 💬
13 / 15
今日之声 VOICE OF THE DAY
AI 工作流的下一步,不是写出更巧的 prompt,而是把人的示范、语音上下文和安全边界沉淀成可复用系统。
The next step for AI workflows is not cleverer prompts, but reusable systems built from human demonstrations, voice context, and safety boundaries.
@claudeaiX 原文37K ❤ · 2.5K RT · 999 💬
14 / 15
AI前沿每日脉动
AI Frontier Pulse · 2026.07.22
本期收录 17 位 Builder · 36 条推文 · 1 期播客 · 1 篇博客
Claude Skills · Voice Context · Agent Security · Work AI · Low-latency Models
感谢阅读 · Thank You For Reading
Richard Liu · AI前沿每日脉动 · 2026 · Snapshot follow-builders-2026-07-22-v1